Improvement of Evolutionary Computation Approaches for Continuous Dynamical System Identification - Robustness and Performance Improvement of Standard Genetic Programming by Approximation, Multiple Shooting Methods, and Iterative Approaches

Created by W.Langdon from gp-bibliography.bib Revision:1.8129

@PhdThesis{Buchsbaum:thesis,

• author = "Thomas Buchsbaum",
• title = "Improvement of Evolutionary Computation Approaches for Continuous Dynamical System Identification - Robustness and Performance Improvement of Standard Genetic Programming by Approximation, Multiple Shooting Methods, and Iterative Approaches",
• school = "Institute of Engineering and Business Informatics, Graz University of Technology",
• year = "2007",
• address = "Kopernikusgasse 24 Graz, Austria",
• keywords = "genetic algorithms, genetic programming, System Identification, Evolutionary Computation, Design of Experiments, Input Signal Shaping, Multiple shooting Approximation Time series modeling, Dynamical systems, Continuous state space models, System Identifikation, Evolutionaere Algorithmen, Zeitreihenanalyse, Versuchsplanung, Eingangsgroessenoptimierung",
• language = "English",
• URL = "https://online.tugraz.at/tug_online/pl/ui/$ctx;lang=DE/wbAbs.showThesis?pThesisNr=24591&pOrgNr=13706",
• URL = "https://graz.elsevierpure.com/en/publications/improvement-of-evolutionary-computation-approaches-for-continuous",
• abstract = "The objective of a mathematical model is to describe certain aspects of a real system; the aim of system identification is to create models from, usually noisy, measurement data. Genetic Programming (GP) is a biology-inspired method for optimising structured representations in general, and dynamical model structures and their parameters in particular. It has been applied to continuous dynamical system identification, but suffered from weak performance and premature convergence behavior. This thesis investigates GP's suitability for creating nonlinear continuous state-space models from noisy time series data. Methodologies are introduced that improve GP's performance and robustness. For the considered test problem, it is shown that instead of solving a dynamical problem by an initial value method, a static problem can be approximated, which can be solved by symbolic regression. This approximation approach speeds up evolution considerably. Fitness evaluation using multiple shooting methods, known from the field of chaotic time series, simplifies the optimization problem by smoothing the fitness landscape; the GP algorithm finds useful building blocks more easily. Three concepts for integrating multiple shooting into GP systems are developed and compared. This thesis offers a concept for automatically switching the identification approach based on the information content of the training data. Computational studies showed that automatic switching combined the advantages of different identification approaches: Better models were created in shorter times. Further, multi-objective methods for regularisation were shown to improve the evolved models generalization abilities substantially. Investigations of model-based input signal optimisation by evolutionary computation methods complete this dissertation. The developed methodologies improve GPs performance and robustness on continuous dynamical system identification tasks. This makes GP a useful tool that assists human modelers in finding building blocks for model synthesis. By applying the introduced methods, the chance of finding hidden information in complex signals can be increased. Medicine, natural sciences, technology, and business could benefit from the improved prediction qualities of the resulting models and the cost savings due to data-efficient modeling procedures.",
• notes = "also known as \cite{711f8ca066c0490eb8e33fadbcd363a9}",

}

Genetic Programming entries for Thomas Buchsbaum

Citations